Development and Applications of a Simulation Framework for a

Wall-Climbing Robot

Daniel Schmidt and Karsten Berns

Abstract— In the range of robotics, simulation frameworksare very common. They are used to perform tests of algo-rithms, for optimization and to analyze the system behavior insituations, which would be hazardous or difficult for the realrobot. This paper addresses a simulation framework relatedto a wall-climbing robot using negative pressure adhesion incombination with an omnidirectional drive system. The keyaspect of this simulation is the adhesion system consisting ofsimulated pressure sensors, valves between adhesion chambersand vacuum reservoir and a simulated adaptive sealing proofingthe vacuum chambers towards ambient air. The interactionof environmental features (e. g. surface characteristics likeroughness or special geometries) and the vacuum chambers ofthe robot is handled by a thermodynamic model providing thebasis for airflow simulation between the virtual surface andthe robot. These features facilitate the validation of controlalgorithms and closed-loop controllers in realtime.

I. INTRODUCTION

Nowadays, robot simulation frameworks are used all over the

world to test algorithms and to analyze the system behavior

in certain situations without executing them on the real

machine. This reduces testing time, avoids unnecessary stress

on the hardware and enables developers to analyze a system

without danger. Current frameworks facilitate a physically

realistic 3D simulation of all kinds of robots from walking

machines up to flying and underwater robots. But so far,

support for the simulation of negative pressure adhesion in

the context of mobile climbing robots is still missing.

This paper presents the simulation of CROMSCI1 which

is a complex wheel-based climbing robot with negative

pressure adhesion (figure 1). The simulation of the inter-

action between robot and environment – especially between

adhesion system and surface – allows extensive tests of safety

strategies and closed-loop controllers without the use of the

real machine. This becomes even more important since real-

world experiments come with a large stress on the hardware

which is not robust enough for long-term tests.

The prototype CROMSCI has been developed to perform

inspection tasks on large concrete buildings area-wide and

semi-autonomously [1]. Its key aspects are seven individ-

ual adhesion chambers generating downforces via negative

pressure, inflatable sliding sealings for leak-tightness, built-

in suction engines evacuating a large vacuum reservoir and

three steerable driven wheels without suspension. It has a

diameter of 80 cm, a height of 40 cm, a weight of about

50 kg, a payload of about 10 kg and a maximal velocity of

D. Schmidt and K. Berns are with the Robotics ResearchLab, University of Kaiserslautern, 67663 Kaiserslautern, Germany{dschmidt,berns}@cs.uni-kl.de

1http://agrosy.cs.uni-kl.de/cromsci/

Fig. 1. Composition of real robot CROMSCI (left) and 3D simulation (right).

9.81 m/min. Its adhesion system generally works in a range

of -50 mBar to -100 mBar compared to ambient pressure to

generate a downforce of about 2 000 N. All these aspects

have been considered for the simulation tool.

Next section II gives a brief overview of existing simu-

lation frameworks. The general concept of the simulation is

presented in section III. Section IV describes the SIMVIS3D

framework used for 3D simulation and rendering. After-

wards, section V presents the thermodynamic model and

leakage simulation. The physical interaction is handled in

section VI. Section VII shows some applications, conclusions

follow in section VIII.

II. RELATED WORK

In robotics, several different simulation tools exist. Most of

these frameworks focus on realistic sensor data and physical

interaction between robot and environment. An example is

Gazebo which supports the well-established robot control

framework ROS [2] and uses ODE or Bullet physic engine.

Also SimRobot is based on ODE and focuses on a realistic

simulation of camera disturbances and actuator friction [3].

USARSim [4] works on the basis of the Unreal game engine

and Karma physics. Webots [5] focuses on research robots

(e. g. Aibo, Pioneer or Kuka youBot). V-REP is a feature-rich

3D simulator for wheel-driven, walking and flying robots

using either ODE or Bullet [6].

Although all these tools are sufficient for collision detec-

tion and pick-and-place tasks, they do not support negative

pressure adhesion as in the case of CROMSCI. Even the

support for particle systems in Vortex and V-REP is not

suitable to simulate airflow since this is a thermodynamic

problem and not related to particle impulses. In the context

of climbing robots there only exist individual solutions for

computing the adhesion forces: [7] uses Simulink to simulate

the drive system of the Alicia3 robot, but only with a

simplified model for the wheel-wall contact and none for

the suction cups. [8] calculates the contact forces between

feet and adhesive objects via continuous mechanics theory

as plugin for Webots.

The main contribution of this paper is the combination

of a thermodynamic model of the negative pressure system,

environmentally conditioned sealing leakages and the drive

system affected by downforce and wheel slippage.

III. CONCEPT

The complete control and simulation software of CROMSCI

has been developed using the MCA framework2. The embed-

ded simulation itself consists of modular components and

connections as depicted in figure 2 an is divided into four

logical blocks shown as gray boxes:

• Actuator simulation simulates ideal actuators for the

omnidirectional drive system, the tool center point, suc-

tion engines and chamber valves for adhesion control.

The desired commands are executed by the actuators

with a certain delay and within individual limits in

velocity, acceleration or volume stream, depending on

the type of actuator.

Fig. 2. Components of the simulation framework (gray).

• Simulation of physical interaction is optional

and can apply physics engines like PhysX or

Newton Game Dynamics with suitable collision and

interaction models. It uses the wheel states (steering

angle and velocity) to calculate wheel slippage and the

resulting robot pose – and finally collisions. Without

physics, the odometry of the robot has to be determined

by hand for subsequent simulation steps.

• SIMVIS3D contains the visual representation of the

scene. It allows it to simulate environmental sensors

(e. g. cameras or laser range sensors) and to visualize the

simulated robot scene. Additionally, SIMVIS3D is used

2http://rrlib.cs.uni-kl.de/mca-kl/

to extract a depth image of the surface below CROMSCI

for sealing simulation.

• Adhesion simulation contains a thermodynamic model

to calculate the progress of airflows and chamber pres-

sures. Therefore, the simulated leakage values as well

as the states of suction engines and valves are used to

calculate an overall adhesion force (robot downforce)

which works contrary to gravity in the physical simu-

lation, if used.

The modular structure of this simulation framework en-

ables the developer to replace single elements by real

hardware components. Therefore, it is e. g. possible to test

the functionality of the real chamber valves via simulated

pressure values. In total, one can enable or disable the real

drive, manipulator, valves, pressure sensors and laser ranger.

IV. 3D SIMULATION AND VISUALIZATION

The simulation framework of CROMSCI uses SIMVIS3D

for simulating and visualizing objects in a 3D scene. This

tool is based on the well-established OpenInventor standard

(OIV) using COIN3D as open source implementation, which

represents a high level abstraction of OpenGL. The core

structure is a scene graph which is set up via a XML

description and arranges the elements in the simulated envi-

ronment and manages scene changes like object positions at

runtime. Scene elements are visual 3D bodies or modifiers

like transformation nodes.

The simulated environment of CROMSCI contains different

components for simulation as well as for decoration. Figure 3

shows the basic wireframe model of the robot and its

surroundings (trees, bridge elements, ground) and a rendered

view including trees, textured surfaces and backgrounds.

SIMVIS3D allows it to move and rotate elements like wheels

and manipulator arbitrarily in the scene and visualize the

result via a graphical user interface (GUI) in realtime to

observe the robot inside the simulation.

Fig. 3. Wireframe model of CROMSCI and a bridge including a band ofrough structured surfaces (top) and rendered visualization (bottom).

The simulated bridge pylon consists of single cubes with

an edge length of 2 m and one structured face. Therefore, it

is possible to insert a cube at any pose via the XML scene file

and to create different test conditions easily. Up to now more

than 50 different cubes with surface structure, cracks, gaps,

holes and protrude structures have been created via Blender

and integrated into the simulation as OpenInventor files.

V. SIMULATION OF NEGATIVE PRESSURE

ADHESION

The most significant components of the simulation frame-

work are those which are responsible for the modeling of

the adhesion system. This includes a calculation of the

airflows between negative pressure chambers and ambient

air as well as a determination of chamber leakages based on

surface characteristics. Figure 4 gives a structural view on the

adhesion system. The top view shows the six outer and one

inner adhesion chambers. The side view below depicts the

mass flows caused by the suction engines mA, valves mVi

and sealing leakages mLij. Beside basic leakages, the leak-

tightness decreases in cases of cracks or surface irregularities.

Fig. 4. Structural view of adhesion chambers 1 to 7, reservoir R and anexemplary crack below chambers 3 and 7.

The interaction model between adhesion system and en-

vironment is based on the first fundamental theorem of

thermodynamics and Bernoulli’s equation as described in [9].

Here, a mass flow mij between two volumes i and j depends

on pressure values, pi and pj , the area of airflow Aij and

the density of air ρair (equation 1):

mij = sgn(pi − pj) ·Aij ·√

2 · ρair · |pi − pj | (1)

Now, the pressure change pi of volume i can be calculated

(equation 2) using the adiabatic index of air κair, the specific

gas constant Rair, temperature Tair and volume Vi. The sum

goes over all other volumes k including the ambient air.

If there is no direct connection between two volumes, the

corresponding massflow is zero.

pi =κair ·Rair · Tair

Vi

∑

k

mik (2)

Given the pressure changes and a starting pressure, the

total affecting downforce F and the point of action ~PF can

be calculated (equations 3 and 4). These equations use the

ambient air pressure po in conjunction with chamber suction

area Ai and chamber center point ~Pi.

F =

7∑

i=1

Fi =

7∑

i=1

((po − pi) ·Ai) (3)

~PF =

7∑

i=1

(

~Pi ·(po − pi) ·Ai

F

)

(4)

The last piece of the puzzle is the calculation of sealing

leakages [10], which are needed to determine e. g. the airflow

between adhesion chamber and ambient air like mLo3in

figure 4. A simplified sealing model is used which approx-

imates its adaptability to surface irregularities. An ortho-

graphic camera (resolution 256×256 pixel) is embedded into

simulation via SIMVIS3D to take a picture (size 80×80 cm2)

of the ground below the robot chassis. The z-buffer of this

camera is used to extract height information. As depicted

in figure 5, the surface height is shown as grayscale image

including an overlay of sealing edges (black and white) and

wheel units. The observed height range is ±5 cm resulting

in a z-resolution of 0.39 mm of the simulated camera.

Fig. 5. Depth image, sealing areas (left) and leakages (right).

In a successional step the adaptability of the sealing to

this height distribution has to be simulated. But, the real

interaction between sealing and surface is to complex to

emulate here. Therefore, an approximation is used which

applies a maximum height difference between neighbor

sealing pixels. Depending on this difference the height values

of all sealing pixels are calculated:

1. determine height of all surface pixels

2. set all sealing pixel heights to surface

. height values below

3. repeat

4. reset changed flag

5. for each sealing pixel

6. if difference to highest neighbor pixel

. larger than maximum height difference

7. set current sealing pixel height to

. highest one minus maximum difference

8. set changed flag

9. until not changed

The result of this algorithm is illustrated in figure 5

(right). Black pixels indicate leak tight areas whereas gray

to white values show permeable areas with a larger gap

between ground and sealing. The total amount of indentation

and extrusion is limited considering the capabilities of the

real robot. In the present case the real sealing can handle

structures of up to ±5mm (so deeper crack cannot be

proofed, higher protrusions cannot be passed). The final

leakage Li (equation 5) of sealing section Si is based on

the length of the sealing segment li, the basic leakage Lbasic

caused by the sliding coating and an multiplication factor κ

for debugging purposes.

Li = li ·

Lbasic + κ ·∑

j∈Si

hs(j)− hg(j)

|Si|

(5)

The sum goes over all pixels j lying on sealing segment

Si and calculates the height difference between sealing

hs(j) and ground pixel hg(j). |Si| is the total amount of

pixels of the sealing segment. Finally, this leakage is used

as connection area between the volumes to calculate the

pressure change pi according to equation 2.

VI. PHYSICAL INTERACTION OF ROBOT AND

ENVIRONMENT

For a realistic system behavior, all actuators are simulated

with certain limitations of acceleration, velocity or position.

These parameters depend on the real hardware settings

and cause e. g. a delayed execution of certain motions like

drive steering or valve positioning. So far, internal hardware

malfunctions (i. e. tolerances, runtime errors or defects) are

of no interest in the present case, but they can be added for

an analysis of their effect on the system.

To complete the simulation framework, a physics engine

can be embedded to simulate the system behavior on af-

fecting forces. Currently, the simulation of CROMSCI uses

Nvidia’s PhysX3, but it could be replaced by other engines

like Newton Game Dynamics4 as it has been done in other

research projects before [11]. In fact, each physic engine

could be used which considers the following aspects:

• Collision models of robot parts and the environment,

• affecting forces in terms of gravity and adhesion force,

• friction model between wheels and ground surface and

• orientation and velocity of wheels for robot locomotion.

Due to simplicity, the physical environment of CROMSCI

does not use the same rough structured mesh as for leakage

simulation. Figure 6 shows the physical mesh model of the

robot consisting of the convex hull of the basic chassis.

Below, the drive system has been emulated via three units,

3http://developer.nvidia.com/physx4http://www.newtondynamics.com

each consisting of one rotational joint for wheel steering and

a PhysX wheel shape which can be turned for locomotion.

Arrows mark the pose of each part which can be annoted

with certain characteristics like dynamic and static friction or

a mass. Parameters and values of each part can be displayed

online, as it is shown on the left side of figure 6.

Fig. 6. Online view on mesh model for physical simulation via PhysX.

The definition of the interface between MCA and PhysX

is done via a configuration class. Here, three different types

of interaction exist which have to be set manually inside of

the code depending on the type of actuator:

• Imports define control values to set e. g. the wheel

speed eIMPORT_WHEEL_SPEED or the adhesion force

eIMPORT_FORCE of the chassis, e. g.if (actor == wheel_actor)

imports |= eIMPORT_WHEEL_SPEED;

if (actor == cromsci_body_actor)

imports |= eIMPORT_FORCE;

• Exports are (virtual) sensor values, mainly used to get

the pose of all objects (eEXPORT_LOCAL_POSE). But

it is also possible to read out other values provided by

the PhysX engine:exports |= eEXPORT_LOCAL_POSE;

if (actor == cromsci_body_actor)

exports |= eEXPORT_ROLL_PITCH_YAW;

• Parameters can be set to adjust global values like

gravity as well as object settings such as mass, fric-

tion or the behavior of the wheels contact point

(ePARAM_WHEEL_TIRE_FUNCTION), e. g.if (actor == wheel_actor)

params |= ePARAM_WHEEL_TIRE_FUNCTION;

The wheel-ground interaction in terms of friction, ve-

locities and affecting forces is simplyfied in two ways. At

first, the surface is assumed to be flat, as mentioned before.

Furthermore, the friction caused by the inflatable sealing is

also not considered here. Nevertheless, one can sufficiently

simulate the robot behavior while the system drives. De-

pending on given parameters like friction values, wheel setup

and generated downforce a different robot behavior relative

to its motion direction and gravity can be observed. For

instance, wheel slip increases due to gravity if the robot

drives upwards. In the same way, the physic engine lets the

robot fall down if its adhesion force is too low.

The current state of the physical simulation can be ob-

served during runtime via a visualization window, as given in

figure 6. It is possible to pause the simulation, walk through

the scene and visualize current values of all components.

Furthermore one can adapt parameters like gravity during

runtime – if they have been registered inside of the simula-

tion’s description – or apply an impact to a simulated body.

VII. APPLICATION

The presented framework has been created to test algorithms

and control strategies which will be applied to the real

prototype CROMSCI. The screenshot of CROMSCI’s GUI in

figure 7 illustrates some parts of the simulation. On the left

side control elements like buttons, sliders and joysticks are

given to set up the closed-loop controllers and activate robot

motion (1). At (2) the chamber states are provided in terms

of a circular robot view showing negative pressure values

as well as some graphical representations of the estimated

leakage or valve opening areas. The current view of the

used 3D scene is given in (3) which can be changed during

runtime to walk through the scene. Below, a plugin visualizes

the current behavior meta values of the control system (4).

On the right side the sealing view (5) and some online plots

of adhesion and downforce data (6) are shown.

12

3

4

5

6

Fig. 7. GUI with visualization plugins for the simulated scene, controlelements and state information.

In total, the simulation framework is used to test and

to develop a couple of aspects related to climbing robots

[1]. This includes software and controllers, but also robot

design in terms of different chamber setups, volume size or

geometric aspects. Without, it would not be possible to bring

CROMSCI this far since its hardware components are still not

robust enough for long-term experiments.

A. Closed-loop Adhesion Control

One example of application is the validation of the individual

closed-loop adhesion controllers of the chambers and subse-

quent safety strategies. Figure 8 (top) shows a situation in

which the robot drives down a rough wall with a deep crack

in front. Here, the left circular view represents the state of

the negative pressure of all seven chambers. The darker the

color, the lower the pressure (the real GUI as given in figure 7

uses colors changing from green to red). The numbers 2, 4

and 6 indicate the associated adhesion chambers, whereas the

dimensions of the black circles in non-labeled chambers 1,

3 and 5 depict the absolute downforce at the corresponding

wheel. It can be observed, that these downforces are balanced

out (all circles have the same diameter) by a higher negative

pressure at the top chambers (darker color of chambers 3, 4

and 5 compared to 1, 2 and 6). This is important to counteract

robot tilt caused by gravity and robot’s mass distribution.

Fig. 8. Screenshots showing the robot driving to a deep crack (top, right)and located on the crack (bottom) with corresponding states of the adhesionsystem (left).

Later on (figure 8, bottom), the lower chambers are located

on the crack and lose adhesion (white chamber color). As a

result the downforces at the lower wheel (circle in chamber

1) are small compared to the two wheels on top.

30.0

15.0

0.0A1[10-5m

2]

30.0

15.0

0.0A4[10-5m

2]

99 000

97 000

95 000

93 000

chamberaffected by

crack

p1[Pa]

99 000

97 000

95 000

93 000

chamberaffected by

reservoir

flat terrain(low

pressurevariance)

rough terrain(more pressure

adjusting needed)

p4[Pa]

100 000

95 000

90 000

85 000

80 000

permanent lossof reservoir

pressure

pR

[Pa]

0 2.5 5 7.5 10 12.5 15 t [s]

Fig. 9. Simulation experiment of closed-loop pressure controllers in whichthe robot tries to overcome a deep crack.

Figure 9 presents corresponding pressure p and valve areas

A of the frontal chamber 1, the rear chamber 4 and the

reservoir pressure pR. At the beginning, the robot is driven

on a very smooth surface with low basic leakages. At about

t = 3.5 s the frontal chamber reaches a structured patch as

given in figure 8. In the following, the chamber controllers

have to perform more pressure balancing and the reservoir

pressure rises slightly. Then, the first chamber 1 reaches

the crack (t = 11 s) and its pressure increases to ambient

pressure very fast. From now on, the controllers try to adjust

the pressure by opening the valves, which results in a loss

of reservoir pressure. The increasing reservoir pressure also

influences those chambers, which are not affected by the

crack directly. This is visualized by the pressure value p4 of

the rear chamber, which also starts to increase up to a value

close to ambient pressure. The valves of all chambers open

to the maximum. In this case the robot has no chance to

maintain the desired adhesion forces and drops off, because

the adhesion system reaches its limits (valves full opened,

suction engines at maximum).

A significant comparison of simulated and real-world

experiments is not easy (the general similarity of the air

flow has been described in [9]). But, one can state that

the simulation is similar enough to transfer the control

algorithms to the high-dynamic machine easily. In the worst

case, some simple parameters need to be adapted.

B. Obstacle Avoidance

Another application of the simulation framework has been

the development and optimization of methods for obstacle

avoidance based on laser range data [12]. Figure 10 shows

a visualization of the simulated scan lines which are emitted

from a simulated sensor on top of the manipulator arm.

Fig. 10. Simulated scan lines of a laser range sensor (left) and resultingscan points (right) as top-view.

Via parameters this simulated sensor can achieve nearly

the same characteristics as the real sensor, such as accuracy,

range and angle. The processing and analysis algorithms as

well as subsequent behavior-based control components have

been tested in the simulation framework first. The simulation

was of great benefit to test these safety features without

endangering the real system. Since the simulation setup was

close enough, the algorithms could be executed on CROMSCI

without any problems, as published in [12].

VIII. CONCLUSIONS

Within the scope of this paper a novel framework has been

presented to simulate a climbing robot equipped with a

negative pressure adhesion system and an omnidirectional

drive system. It could be shown that the system can simulate

aspects like airflows and pressure changes of the complex

vacuum system as well as environmental sensors for obstacle

detection. Related to the negative pressure adhesion, the

surface structure has a direct impact on the sealing leakages

to simulate different ground characteristics like roughness or

defects. Hence, this tool supports the development of control

algorithms independent of the real robot, enables an offline

and realtime validation in arbitrary 3D environments and

allows an easy testing of different robot designs and setups.

In future, this simulation will be ported to the new robot

control framework FINROC5. Further work lies in an easier

handling of the PhysX integration, since the Blender export

of collision objects does not include joints (they have to be

added manually so far). Finally, the new robot prototype

CREA is under construction. Its control software – closed-

loop controllers as well as high deliberative functions –

will be developed and optimized within the simulation first,

before executing it on the real machine.

ACKNOWLEDGEMENTS

This research was funded by the German Bundesministerium

fuer Wirtschaft und Technologie (BMWi) under the program

Zentrales Innovationsprogramm Mittelstand (ZIM).

REFERENCES

[1] D. Schmidt, “Safe Navigation of a Wall-Climbing Robot - RiskAssessment and Control Methods,” doctoral thesis, University ofKaiserslautern, Germany, 2013.

[2] N. Koenig and A. Howard, “Design and use paradigms for gazebo,an open-source multi-robot simulator,” in International Conference on

Intelligent Robots and Systems (IROS), 2004.[3] T. Laue and T. Rofer, “Simrobot - development and applications,” in

International Conference on Simulation, Modeling and Programming

for Autonomous Robots (SIMPAR), 2008.[4] J. Wang, M. Lewis, and J. Gennari, “A game engine based simulation

of the nist urban search and rescue arenas,” in Winter Simulation

Conference, 2003.[5] O. Michel, “Webots: Professional mobile robot simulation,” Interna-

tional Journal Of Advanced Robotic Systems, vol. 1, no. 1, 2004.[6] M. Freese, S. Singh, F. Ozaki, and N. Matsuhira, “Virtual robot

experimentation platform v-rep: A versatile 3d robot simulator.” in 2nd

International Conference on Simulation, Modeling and Programming

for Autonomous Robots (SIMPAR), 2010.[7] D. Longo, G. Muscato, and S. Sessa, “Simulation and locomotion

control for the alicia3 climbing robot,” in International Symposium on

Automation and Robotics in Construction (ISARC), 2005.[8] I. Pretto, S. Ruffieux, C. Menon, A. Ijspeert, and S. Cocuzza, “A

point-wise model of adhesion suitable for real-time applications ofbio-inspired climbing robots,” Journal of Bionic Engineering, vol. 5,2008.

[9] J. Wettach, C. Hillenbrand, and K. Berns, “Thermodynamical mod-elling and control of an adhesion system for a climbing robot,” inInternational Conference on Robotics and Automation (ICRA), 2005.

[10] D. Schmidt, J. Wettach, and K. Berns, “3d realtime simulation frame-work for a wall-climbing robot using negative-pressure adhesion,” in10th International Conference on Informatics in Control, Automation

and Robotics (ICINCO), July 29-31 2013.[11] J. Wettach, D. Schmidt, and K. Berns, “Simulating vehicle kinematics

with simvis3d and newton,” in 2nd International Conference on Simu-

lation, Modeling and Programming for Autonomous Robots (SIMPAR),2010.

[12] M. Jung, D. Schmidt, and K. Berns, “Behavior-based obstacle de-tection and avoidance system for the omnidirectional climbing robotcromsci,” in 13th International Conference on Climbing and Walking

Robots (CLAWAR), August 31 - September 3 2010, pp. 73–80.

5http://www.finroc.org